Hi Alison,
On 2/10/2026 11:16 AM, Alison Schofield wrote:
On Tue, Feb 10, 2026 at 06:44:52AM +0000, Smita Koralahalli wrote:
This series aims to address long-standing conflicts between HMEM and
CXL when handling Soft Reserved memory ranges.
Reworked from Dan's patch:
https://git.kernel.org/pub/scm/linux/kernel/git/cxl/cxl.git/patch/?id=ab70c6227ee6165a562c215d9dcb4a1c55620d5d
Previous work:
https://lore.kernel.org/all/[email protected]/
Link to v5:
https://lore.kernel.org/all/[email protected]
The series is based on branch "for-7.0/cxl-init" and base-commit is
base-commit: bc62f5b308cbdedf29132fe96e9d591e526527e1
[1] After offlining the memory I can tear down the regions and recreate
them back. dax_cxl creates dax devices and onlines memory.
850000000-284fffffff : CXL Window 0
850000000-284fffffff : region0
850000000-284fffffff : dax0.0
850000000-284fffffff : System RAM (kmem)
[2] With CONFIG_CXL_REGION disabled, all the resources are handled by
HMEM. Soft Reserved range shows up in /proc/iomem, no regions come up
and dax devices are created from HMEM.
850000000-284fffffff : CXL Window 0
850000000-284fffffff : Soft Reserved
850000000-284fffffff : dax0.0
850000000-284fffffff : System RAM (kmem)
[3] Region assembly failure works same as [2].
[4] REGISTER path:
When CXL_BUS = y (with CXL_ACPI, CXL_PCI, CXL_PORT, CXL_MEM = y),
the dax_cxl driver is probed and completes initialization before dax_hmem
probes. This scenario was tested with CXL = y, DAX_CXL = m and
DAX_HMEM = m. To validate the REGISTER path, I forced REGISTER even in
cases where SR completely overlaps the CXL region as I did not have access
to a system where the CXL region range is smaller than the SR range.
850000000-284fffffff : Soft Reserved
850000000-284fffffff : CXL Window 0
850000000-280fffffff : region0
850000000-284fffffff : dax0.0
850000000-284fffffff : System RAM (kmem)
"path":"\/platform\/ACPI0017:00\/root0\/decoder0.0\/region0\/dax_region0",
"id":0,
"size":"128.00 GiB (137.44 GB)",
"align":2097152
[ 35.961707] cxl-dax: cxl_dax_region_init()
[ 35.961713] cxl-dax: registering driver.
[ 35.961715] cxl-dax: dax_hmem work flushed.
[ 35.961754] alloc_dev_dax_range: dax0.0: alloc range[0]:
0x000000850000000:0x000000284fffffff
[ 35.976622] hmem: hmem_platform probe started.
[ 35.980821] cxl_bus_probe: cxl_dax_region dax_region0: probe: 0
[ 36.819566] hmem_platform hmem_platform.0: Soft Reserved not fully
contained in CXL; using HMEM
[ 36.819569] hmem_register_device: hmem_platform hmem_platform.0:
registering CXL range: [mem 0x850000000-0x284fffffff flags 0x80000200]
[ 36.934156] alloc_dax_region: hmem hmem.6: dax_region resource conflict
for [mem 0x850000000-0x284fffffff]
[ 36.989310] hmem hmem.6: probe with driver hmem failed with error -12
[5] When CXL_BUS = m (with CXL_ACPI, CXL_PCI, CXL_PORT, CXL_MEM = m),
DAX_CXL = m and DAX_HMEM = y the results are as expected. To validate the
REGISTER path, I forced REGISTER even in cases where SR completely
overlaps the CXL region as I did not have access to a system where the
CXL region range is smaller than the SR range.
850000000-284fffffff : Soft Reserved
850000000-284fffffff : CXL Window 0
850000000-280fffffff : region0
850000000-284fffffff : dax6.0
850000000-284fffffff : System RAM (kmem)
"path":"\/platform\/hmem.6",
"id":6,
"size":"128.00 GiB (137.44 GB)",
"align":2097152
[ 30.897665] devm_cxl_add_dax_region: cxl_region region0: region0:
register dax_region0
[ 30.921015] hmem: hmem_platform probe started.
[ 31.017946] hmem_platform hmem_platform.0: Soft Reserved not fully
contained in CXL; using HMEM
[ 31.056310] alloc_dev_dax_range: dax6.0: alloc range[0]:
0x0000000850000000:0x000000284fffffff
[ 34.781516] cxl-dax: cxl_dax_region_init()
[ 34.781522] cxl-dax: registering driver.
[ 34.781523] cxl-dax: dax_hmem work flushed.
[ 34.781549] alloc_dax_region: cxl_dax_region dax_region0: dax_region
resource conflict for [mem 0x850000000-0x284fffffff]
[ 34.781552] cxl_bus_probe: cxl_dax_region dax_region0: probe: -12
[ 34.781554] cxl_dax_region dax_region0: probe with driver cxl_dax_region
failed with error -12
v6 updates:
- Patch 1-3 no changes.
- New Patches 4-5.
- (void *)res -> res.
- cxl_region_contains_soft_reserve -> region_contains_soft_reserve.
- New file include/cxl/cxl.h
- Introduced singleton workqueue.
- hmem to queue the work and cxl to flush.
- cxl_contains_soft_reserve() -> soft_reserve_has_cxl_match().
- Included descriptions for dax_cxl_mode.
- kzalloc -> kmalloc in add_soft_reserve_into_iomem()
- dax_cxl_mode is exported to CXL.
- Introduced hmem_register_cxl_device() for walking only CXL
intersected SR ranges the second time.
During v5 review of this patch:
[PATCH v5 6/7] dax/hmem, cxl: Defer and resolve ownership of Soft Reserved
memory ranges
there was discussion around handling region teardown. It's not mentioned
in the changelog, and the teardown is completely removed from the patch.
The discussion seemed to be leaning towards not tearing down 'all', but
it's not clear to me that we decided not to tear down anything - which
this update now does.
And, as you may be guessing, I'm seeing disabled regions with DAX children
and figuring out what can be done with them.
Can you explain the new approach so I can test against that intention?
FYI - I am able to confirm the dax regions are back for no-soft-reserved
case, and my basic hotplug flow works with v6.
-- Alison
Hi Alison,
Thanks for the test and confirming the no-soft-reserved and hotplug
cases work.
You're right that cxl_region_teardown_all() was removed in v6. I should
have called this out more clearly in the changelog. Here's what I learnt
from v5 review. Correct me if I misunderstood.
During v5 review, regarding dropping teardown (comments from Dan):
"If we go with the alloc_dax_region() observation in my other mail it
means that the HPA space will already be claimed and
cxl_dax_region_probe() will fail. If we can get to that point of "all
HMEM registered, and all CXL regions failing to attach their
cxl_dax_region devices" that is a good stopping point. Then can decide
if a follow-on patch is needed to cleanup that state
(cxl_region_teardown_all()) , or if it can just idle that way in the
messy state and wait for userspace to cleanup if it wants."
https://lore.kernel.org/all/[email protected]/
Also:
"In other words, I thought total teardown would be simpler, but as the
feedback keeps coming in, I think that brings a different set of
complexity. So just inject failures for dax_cxl to trip over and then we
can go further later to effect total teardown if that proves to not be
enough."
https://lore.kernel.org/all/[email protected]/
The v6 approach replaces teardown with the alloc_dax_region() resource
exclusion in patch 5. When HMEM wins the ownership decision (REGISTER
path), it successfully claims the dax_region resource range first. When
dax_cxl later tries to probe, its alloc_dax_region() call hits a
resource conflict and fails, leaving the cxl_dax_region device in a
disabled state.
(There is a separate ordering issue when CXL is built-in and HMEM is a
module, where dax_cxl may claim the dax_region first as observed in
experiments [4] and [5], but that is an independent topic and might not
be relevant here.)
So the disabled regions with DAX children you are seeing on the CXL side
are likely expected as Dan mentioned - they show that CXL tried to claim
the range but HMEM got there first. Though the cxl region remains
committed, no dax_region gets created for it because the HPA space is
already taken.
Thanks
Smita
